Collective cognition and decision-making in humans and fish

(1)

Collective cognition and decision-making in humans and fish

Dissertation

zur Erlangung des akademischen Grades Doctor rerum agriculturarum (Dr. rer. agr.)

eingereicht an der Lebenswissenschaftlichen Fakultät der Humboldt-Universität zu Berlin

von

Romain Jean Gilbert Clément, M. Sc.

Präsident der Humboldt-Universität zu Berlin: Prof. Dr. Jan-Hendrik Olbertz Dekan der Lebenswissenschaftlichen Fakultät: Prof. Dr. Richard Lucius

Gutachter: 1. Prof. Dr. Jens Krause 2. Dr. Max Wolf

3. Dr. Richard James

Tag der mündlichen Prüfung: 12. April 2016

(2)

(3)

1

Summary

Group living is a widespread phenomenon. One of its assumed advantages is collective cognition, the ability of groups to solve cognitive problems that are beyond single individuals’ abilities. In this thesis, I investigated whether decision-making improves with group size in both humans and fish, thus using the strengths of each system.

In humans, I tested individual performance in simple quantity estimation tasks and a more difficult sentence reconstruction task first alone and then as part of a group. My question was whether groups were able to improve not only on average individual decisions, but also to beat their best members. Indeed, when a given problem is recurrent or too complex for individuals, groups were able to outperform their best members in different contexts. Furthermore, I showed that in a simulated predation experiment, groups of humans decided to stay or to escape using quorum thresholds based on movement behaviour without verbal communication, as has been shown in other animals.

This simple movement mechanism allowed individuals in groups to simultaneously increase true positives and decrease false positives.

In the guppy, a freshwater fish from Trinidad, I tested in their natural environment whether individuals’ ability to distinguish between an edible and a non-edible food item increases with group size. My results indicate that guppies had better chances to identify the edible food item when part of bigger groups. By investigating several populations with different ecological backgrounds, in particular differing in predation levels, I found that, despite a lower sampling activity in high predation habitats, predation did not affect the improvement of decisions in groups.

Overall, this thesis contributes to an improved understanding of collective decision-making, showing that collective cognition can arise from various interaction rules, such as simple aggregation of individual estimates, visual observation of others’

movements, and group discussion. Furthermore, having tested problems ranging from estimating a quantity, distinguishing between edible and non-edible food items, deciding to stay or escape, and reconstructing a distorted message, I showed that grouping with others is beneficial in many situations for both humans and fish. This may suggest that the underlying mechanisms of collective cognition are remarkably similar across species.

(6)

4

Zusammenfassung

Das Zusammenleben in Gruppen ist im Tierreich ein weit verbreitetes Phänomen. Einer der Vorteile des Gruppenlebens könnte die sogenannte „Schwarmintelligenz“ sein, das heißt die Fähigkeit von Gruppen kognitive Probleme zu lösen, die die Problemlösekompetenz einzelner Individuen übersteigt. In der vorliegenden Dissertation untersuchte ich, ob die Gruppengröße beim Menschen und bei Fischen mit einer verbesserten Entscheidungsfindung einhergeht.

Beim Menschen analysierte ich zunächst das Abschneiden von Einzelpersonen, die später als Teil einer Gruppe getestet wurden, in einfachen Einschätzungsaufgaben sowie komplizierteren Satz-Rekonstruktionstests. Meine Frage war, ob es Individuen in Gruppen gelingt bessere Entscheidungen zutreffen als das einem durchschnittlichen Individuum der Gruppe alleine möglich wäre und ob Gruppen sogar die Leistung ihres besten Mitglieds in den individuellen Tests überbieten könnten. Tatsächlich konnte ich zeigen, dass Gruppen die Leistung des besten Mitglieds übertreffen, wenn die Problemstellung für Einzelpersonen zu komplex ist oder sich häufig wiederholt.

Weiterhin gelang mir zu zeigen, dass Gruppen von Menschen bei einer simulierten Prädationssituation, ähnlich wie es bereits für andere Tierarten beschrieben wurde, anhand von so genannten „Quorum“-Regeln durch non-verbale Kommunikation entscheiden, ob sie bleiben oder flüchten. Dabei dienen einfache Bewegungsmuster als Schlüsselreiz. Individuen einer Gruppe erhöhen durch diesen Mechanismus gleichzeitig ihre echt positiven und verringern ihre falsch positiven Entscheidungen.

Beim Guppy, einem Süßwasserfisch aus Trinidad, untersuchte ich in deren natürlichem Habitat, ob die Fähigkeit einzelner Individuen zwischen einer genießbaren und einer ungenießbaren Futterquelle zu unterscheiden, mit der Gruppengröße ansteigt.

Meine Ergebnisse zeigen, dass Guppys mit größerer Wahrscheinlichkeit eine genießbare Futterquelle identifizierten, sobald sie Teil einer größeren Gruppe waren. Untersuchungen an verschiedenen Populationen, die sich vor allem bezüglich des jeweiligen Prädationsdrucks in ihrem Habitat unterschieden, ergaben weiterhin, dass sich, abgesehen von einer niedrigeren Sampling-Rate in Habitaten mit hohem Prädationsdruck, Prädation nicht auf die Qualität der Gruppenentscheidungen auswirkt.

Die vorliegende Arbeit trägt zu einem tieferen Verständnis von kollektiver Entscheidungsfindung bei. Kollektive Intelligenz entsteht aus verschiedensten

(7)

5

Interaktionsregeln, beispielsweise durch die simple Zusammenfassung individueller Vorhersagen, visuelle Observation der Bewegung anderer und Gruppendiskussionen.

Meine Arbeit lässt den Schluss zu, dass bei Problemen wie der Schätzung einer Menge, der Unterscheidung zwischen einer genießbaren von einer ungenießbaren Futterquelle, der Entscheidung zum Bleiben oder Flüchten und der Rekonstruktion einer Nachricht, die Bildung einer Gruppe, sowohl bei Menschen als auch bei Fischen, von Vorteil ist. Diese auffallenden Ähnlichkeiten über Artgrenzen hinweg weisen möglicherweise auf sehr universelle Mechanismen kollektiver Intelligenz hin.

(8)

6

Acknowledgments

I am grateful to the Leibniz-Insitute for Freshwater Ecology and Inland Fisheries (IGB) that provided me with space and funding during my PhD. IGB was an excellent and stimulating place to do my PhD. I also thank the Association for the Study of Animal Behaviour for travel grants to conferences.

I thank my supervisor Jens Krause, who shared his enthusiasm for science and wildlife, and who taught me many lessons, in experimental design, scientific reasoning and writing, und alles. Additionally, he has been an excellent travel companion during field work under conditions that were not always easy.

I thank Ralf Kurvers for the many discussions during which I learnt a lot, in particular concerning advanced statistical analysis, and also for his always swift and thorough feedback on my writings. Ralf’s contribution is indeed reflected in that he is a coauthor on most of the chapters in the thesis. I am also grateful to Max Wolf for his insightful contributions during conversations and on manuscripts.

I am indebted to Stefan Krause for his crucial role during the analysis of complicated datasets involving language processing and other advanced computational methods. Discussions with Dick James, Kate Laskowski, Richard Mann, Thomas Mehner, James Herbert-Read and Ashley Ward improved various parts of this thesis in many ways. I am indebted to David Bierbach for his helpful feedback on the summary and the German translation. I thank Alex Wilson for his sound advice during the early stages of my PhD, and Karoline Borner for regular exchanges during our PhD years.

I am grateful to Indar Ramnarine who greatly facilitated our work in Trinidad. I also want to thank Nikolaus von Engelhardt, Leif Engqvist and Fritz Trillmich for hosting the experiments in Bielefeld, and all the participants of the 2011 and 2012 student course

“Basismodul Biologie” and its tutors. I thank Marcus Ebert, Philipp Beer, Knut Hinrichs, Dominik Jost, Ivan Rodriguez-Pinto, Sonja Smith, and Simon Stäblein for assistance with data collection.

Finally, I thank my family and friends at IGB for their support during the long process of writing a PhD thesis. I thank Carolina for her support during the last months of the PhD. Last, I thank Newton, quiet observer of my writing who helped me focussing during the last phase.

(9)

7

General introduction

1.1 Collective behaviour

Collective behaviour is a widespread phenomenon that has long puzzled observers.

Despite costs such as competition and increased probability of disease transmission, the ubiquity of group living across many species reflects the many advantages that it provides. Such advantages include antipredator defence, finding a mate, conserving heat and water, and reducing the energetic costs of movement (Krause and Ruxton, 2002;

Sumpter, 2010).

At the beginning of the twentieth century, the naturalist Edmund Selous invoked telepathic faculties to explain the ability of flocks to perform fast synchronous changes of directions (Couzin, 2009). Now, various models have been developed to help explain the underlying mechanisms that allow fish schools and bird flocks to execute the synchronized fast group movements and escape manoeuvres. Early models based on particle physics used simple rules such as attraction to other individuals, alignment with neighbours, and repulsion from individuals that are too close (Breder, 1954; Radakov, 1973). Some other models known as topological models are based on a fixed number of neighbours that interact with the focal individual, regardless of their distance, rather than the number of individuals present within a given distance of the focal individual (Ballerini et al., 2008). More recently appeared visual models (Strandburg-Peshkin et al., 2013), in which individuals interact with other individuals that appear in their visual field (and occupy an angular area on their retina that is bigger than a threshold value). These recent models account for a great flexibility allowing groups to be very compact or very loose while maintaining cohesion and effective information transfer between the members. Understanding how members interact in a group is important to appreciate how information flows between members and enables group decision-making.

Moving from the interaction mechanisms governing groups, another important body of literature deals with understanding the adaptive consequences of group living.

The dilution effect and the confusion effect, which are respectively a decrease in the probability to get caught (proportional with the number of individuals in the group) and an increase in the difficulty for the predator to focus on a particular prey in the group

(10)

8

(Ioannou et al., 2012) contribute to a better protection against predators for individuals in groups.

Some advantages come from cooperation and division of labour, specialized individuals being more effective at their task as has been shown in group-hunting dolphins (Gazda et al., 2005), or more commonly known in social insects.

In addition to group advantages that are mostly due to an increase in physical power, being part of large groups can provide advantages at the cognitive level, enabling individuals in groups to make better decisions, such as choosing the best navigation route, or detecting a predator earlier and from a greater distance through collective vigilance, known as the many-eyes effect (Pulliam, 1973), as observed in taxa ranging from water skaters (Treherne and Foster, 1980) to birds (Kenward, 1978).

1.2 Collective decision-making

All organisms, throughout their lives, are faced with many decisions. Whether to forage alone or in a group? What to eat? When and where to go? Whom to mate with? Where to breed? Etc. Therefore a lot of animal behaviour research investigates questions related to choice and decision-making. Decision-making in humans regarding every aspect of their life has been contemplated by philosophers since the antiquity.

Collective decision-making in humans

Owing to the inherent sociality of humans and the importance of decision-making in domains such as politics, economics and law, collective decision-making has been widely studied in humans. Building on pioneering works by Condorcet (1785) and Galton (1907), group decision-making has been studied extensively in humans by psychologists and economists (Fernandez-Juricic et al., 2004; Kerr and Tindale, 2004; Laughlin et al., 2002, 2003; Sunstein, 2005, 2005; Surowiecki, 2004).

However, collective decision-making has been shown to have limits. Depending on the conditions, group decision can either increase or decrease the quality of the decision (Esser, 1998; Janis, 1971; Koriat, 2012; Lassila, 2008). Social influence for example, by reducing independence, can be detrimental as the bare knowledge of others’

opinion may negatively affect the collective outcome, even without communication (Lorenz et al., 2011).

(11)

9

Recently, collective decision-making in humans has been studied a lot in behavioural research with an evolutionary or behavioural ecology focus (Dyer et al., 2008; King et al., 2011a, 2011b; Krause et al., 2010, 2011b; Kurvers et al., 2014a; Wolf et al., 2013).

Collective decision-making in (non-human) animals

Despite the numerous studies on group decision-making in humans and individual decision-making in animals, collective decision-making has long been neglected in animals and the number of studies investigating collective decision-making in non-human animals increased sharply only recently.

In many situations, conflicts of interest can arise (for instance, the optimal choice for an individual can vary within a group depending on its age, sex, nutritional state, etc).

In order to maintain cohesion, animals that live in groups need to make consensual decisions. But, except from social insects, most animal groups are heterogeneous and it is unlikely that all individuals have the same needs at the same time (Conradt, 2011, 2012;

Conradt and List, 2009, 2009; Conradt and Roper, 2003, 2005, 2007, 2009, 2010; Conradt et al., 2009).

Collective decision-making can result in benefits (increased vigilance through many eyes, faster decision, increased navigational accuracy, self-organized orientation to local neighbourhood) but can also present several costs (increased decision time, competition, disease transmission, cascade of information). This thesis deals with situations where it is in the interest of all members to take the best decision, such as escaping from a predator, or deciphering a message. Only chapter 3 may involve some level of competition between the members of a group, but all have the same interest in making the correct choice.

Collective cognition

Groups are able to achieve things that individuals cannot and improved decision-making is at the root of several of the advantages offered by group-living. For example, water skaters, through the many eyes effect, are able to detect predators from a greater distance.

The many eyes effect is an example of collective cognition and leads to an increase of overall vigilance for the group, while allowing a decrease of vigilance at the individual level (Treherne and Foster, 1980).

(12)

10

Collective cognition is a particular case of collective behaviour that has its origins in the fields of complexity and self-organization. It is defined as a process by which “two or more individuals independently collect information that is processed through social interaction and provides a solution to a cognitive problem that is not available to single individuals” (Krause et al., 2010). In short, collective cognition is the ability that groups have to solve cognitive problems that are beyond individual capabilities. Depending on the field in which it is used, or the organisms studied, collective cognition is also known as collective intelligence when used generally on animals (Couzin, 2009), swarm intelligence when applied more specifically to social insects or algorithms (Bonabeau et al., 1999), the Wisdom of Crowds (Surowiecki, 2004), group decision-making or team decision-making in psychology (Kerr and Tindale, 2004; Sunstein, 2005). The expression Swarm Intelligence was first coined by Beni & Wang (1989), to describe “systems of non-intelligent robots exhibiting collectively intelligent behaviour evident in the ability to unpredictably produce ‘specific’ (i.e. not in a statistical sense) ordered patterns of matter in the external environment”. The term “collective intelligence” was first applied to biological systems by Franks (1989) to describe the ability of ant colonies to solve specific problems that appear out of reach for individual ants. It has since then been applied not only to other social animals, ranging from bees (Garnier et al., 2007) to humans (Krause et al., 2011b), and has even been proposed as a coordination mechanism in complex plants roots systems (Baluska et al., 2010).

Indeed, it has been shown that ants are able to choose the shortest route between a food source and their nest using trail pheromone (Goss et al., 1989). Using simple quorum rules, accurate decisions are achieved by bees (Seeley et al., 2006; Visscher and Camazine, 1999) and ant colonies (Pratt et al., 2002) that are forced to find a new nest after being evicted. It has also been shown that speed and accuracy augment with group size in fish shoals avoiding a predator model (Sumpter et al., 2008a; Ward et al., 2008, 2011)

Collective cognition can be achieved through different mechanisms. In addition to benefiting from the possibility to follow a clear leader (Couzin et al., 2005; Danchin et al., 2004; Reebs, 2000), being part of the group can also improve decision-making even when nobody has clear information (Couzin et al., 2011). For example it has been demonstrated that migrating groups use averaging in order to navigate more accurately and reach their destination (Codling et al., 2007; Faria et al., 2009; Hancock et al., 2006;

Wallraff, 1978, 2001). Additionally, group members are still able to distrust a single bad

(13)

11

leader leading them close to a dangerous area and need a critical number (Ward et al., 2011).

In a group, information is rarely equally distributed among all individuals.

Individuals may, for example, differ in their proximity to a predator, or in previous experience about a food location or a past migration route. Therefore leadership can arise in non-hierarchical groups from differences in information. Couzin et al. (2005) predicted that a minority of informed members as small as 5 to 10 % is necessary to lead a whole group towards the right location. This has been empirically verified by Dyer et al. (2008, 2009) using human volunteers. Moreover, recent studies have shown that the presence of uninformed individuals in a group not only favours majority decisions by counterbalancing despotic tendencies of minorities (Couzin et al., 2011), but also improve the stability of the decision-making (Leonard et al., 2011).

1.3 Outline of the thesis

The main aim of my thesis was to investigate whether decision-making improves with group size. For this, I compared individual performance with group performance using an array of simulation experiments on humans (chapters 2, 4, 5) and a field study on Trinidadian guppies, Poecilia reticulata (chapter 3). Using two different systems allowed me to investigate collective cognition from different angles, each system presenting its own strengths.

While investigating whether group decisions improve compared to decisions of individuals, I was also interested in finding out whether collective decisions can also be better than decisions of the best individuals. This is important in order to understand whether top performers join groups because of cognitive benefits or because of benefits related to other group advantages such as, for example, dilution effect. I tackled this question in chapter 2 where I studied groups’ potential performance at solving a recurring task, and in chapter 5, where I studied groups’ performance at solving a more complex task. Studying humans allowed me to know someone’s private information before they acted on it, and before they shared it with other members of the group. This allowed delineating private information from social information, which can be difficult and time-consuming in animals (because this requires individual testing; therefore, most animal studies on collective cognition compared individuals with groups made of different individuals that have not been tested as singletons, making it difficult to assess

(14)

12

to which extent particular individuals improve when part of a group). Still taking advantage of the possibility to know individuals’ decision before and after interaction with others, in chapter 4 I studied the ability of groups to improve decisions using a design that simulated natural situations and that allowed individuals to interact via movement, mimicking the decision process observed in wild animal groups.

Finally, I wanted to see whether the cognitive benefits observed in human groups apply to animals living in natural conditions. Most fundamentally, I was interested in exploring ecological selective pressures that may have influenced the evolution of collective cognition in wild animals. This is possible by comparing different species or different populations and trying to understand how differences in behaviour reflect differences in ecology (Davies et al., 2012). In chapter 3, I used the comparative method to investigate the effect of predation on collective cognition in Trinidadian guppies. I compared the decision improvements in groups across four populations that differed in predation pressure, i.e. two with a high predation level, and two with a low predation level. Furthermore, in this experiment, the group sizes used occurred naturally and were not artificially made by separating already existing groups or putting together unfamiliar individuals.

I chose to present the chapters of my thesis following the increasing complexity of the interactions between group members, which also reflects the increasing level of complexity of the tasks that they had to solve.

Chapters 2 and 3 used a simple assessment of collective cognition based on decision accuracy.

It is known that groups generally outperform average individuals at estimation tasks. But they are often beaten by one or a few individuals whose estimates are even closer to the real value. Chapter 2 explored the advantage of collective cognition compared to top performing individuals in solving a task that is cognitively simple, but repeatedly encountered over time. In this study, individuals were asked to estimate the number of dots that appear on a screen. This estimation task was carried out repeatedly to find out how often individuals beat the group. I compared performance of the best individuals to group performance obtained from a combination of the estimates given by individual members. Depending on the problem that is encountered, the incentive to join the group may be strong enough even for the best individual. For example, in the case of

(15)

13

predator detection, missing to spot a predator only once can be fatal and joining the group is beneficial if such a problem is encountered frequently.

Chapter 3 extended collective cognition to animal groups in the wild. Testing the performance of individuals in various group sizes, I explored the ability of guppies to distinguish two stimuli (one edible and the other not) that are visually very similar. This experiment was run across several populations to investigate the effect of predation pressure on decision-making abilities, and in particular differences in the use of private and social information.

Chapters 4 and 5 took advantage of humans as a study system to investigate collective cognition to solve more complex tasks and a possibility to look into the collective decision mechanism at a finer level.

In chapter 4, groups of humans had to repeatedly distinguish between two cryptic images in a simulated predation detection experiment in which they were only allowed to indicate their preference via movement, in a way that is similar to group decision-making observed in some wild animal groups (such as in mammal herds, bird flocks, fish schools).

It is generally difficult to investigate complex problems as the correct solutions are not always obvious and most of the previous work on swarm intelligence has been carried out on simple estimation tasks. In chapter 5, I explored collective cognition as a tool for solving more complicated problems. After listening to a distorted and hardly understandable announcement, participants were asked to reconstruct the original message individually or as a group, by discussing.

(16)

14

(17)

15

Collecti e cognition otential in umans Grou s can out erform ig - erforming indi iduals

Clément RJG, Krause S, Faria JJ & Krause J

The possibility that individuals in animal groups, including humans, can make collective decisions that solve cognitive problems that single individuals cannot solve, or not in the same way, has attracted much attention in recent years in the context of collective cognition (CC). A common problem when comparing the problem-solving ability of groups and single individuals is that many studies only looked at one-off performances which makes it difficult to distinguish between individuals of high ability and those that made a lucky guess. Here we examined performance profiles of individuals regarding a repeated quantity estimation task which demonstrated that there was significant variation in cognitive ability within a human population. For a single estimate, 13.4% of individuals could beat the group performance. However, if repeated estimates were taken into account, then group performance was superior to even the best individual performers in the group after a sequence of at least 8 estimates. This result suggests that for certain cognitive problems that are encountered repeatedly, joining groups may always be advantageous, even for individuals of high cognitive ability. We discuss our results in the context of the evolution of collective cognition.

Unpublished manuscript

(18)

16

Introduction

One of the benefits of group living in animals comes from improved decision-making resulting from independent information collection and processing among group members (Krause et al., 2010). The fact that individuals in groups can solve cognitive problems in a way that cannot be implemented by single individuals – a process known as collective cognition – has attracted a large amount of interest in several areas (biology: Couzin, 2009; Krause et al., 2010, psychology: Kerr and Tindale, 2004; Sunstein, 2005;

sociology: Mathieu et al., 2008; economics: Armstrong, 2001; Wolfers and Zitzewitz, 2004).

In humans, it has been shown that, for many kinds of problems, the group’s collective performance is better than that of an average individual’s (Hong and Page, 2004; Sunstein, 2005). This phenomenon, known as the “wisdom of crowds” or “many- wrongs principle” was pointed out as early as 1907 by Galton and seems to play an important role in some group formations such as migrating animals (Bergman and Donner, 1964; Hamilton, 1967; Simons, 2004; Wallraff, 1978) and several models support this hypothesis (Codling et al., 2007; Grünbaum, 1998; Hancock et al., 2006).

These models are confirmed by Dell’Ariccia et al. (2008), who showed that homing pigeons travel more efficiently in groups than alone and reach their destination faster.

However, Biro et al. (2006) showed that if the disagreement regarding the route is too high, either the group would split or one pigeon would become the leader. In humans, when the information is unevenly distributed among the members of a group, knowledgeable individuals can still lead the group to the right destination, even without communication between members (Dyer et al., 2008, 2009). But when nobody owns enough information, the wisdom of crowds plays an important role in reaching the correct destination when the group size is large enough and when uncertainty is high (Faria et al., 2009). Dyer et al. (2008, 2009) showed that even when the information is not evenly distributed and only a minority is able to lead, groups of humans still reach their destination, even without communication. Some individuals tend to overestimate while others underestimate and groups usually outperform individuals because averaging these estimations leads to a reduction of the error component (Fischer and Harvey, 1999;

Simons, 2004). When an objective and demonstrable answer does exist such as estimating the temperature of a room, the number of beans in a jar or the ranking of different weights (Sunstein, 2005 and references therein), the group’s mean or median answer often comes

(19)

17

very close to the real value and is better than the vast majority of individual estimates.

However, it has been pointed out that the group performance usually did not surpass the best individual in a given group and that group processing can actually lead to gains as well as losses (Kerr and Tindale, 2004). Very few studies actually reported groups’

outcomes that were better than the performance of every member (Kerr and Tindale, 2004). It is generally considered a truism that groups do better than single individuals, since most statistics is based on the assumption that a larger sample size is more likely to return an average value closer to the true mean. But this is not always the case. For example, Krause et al. (2011) showed that in some cases, as sample size increases, the mean becomes worse, and when it comes to decision-making, benefits from grouping behaviour are not always high enough to outcompete the best individuals.

Furthermore, interactions between members of the group do not always improve the group’s outcome because of inherent obstacles such as peer pressure or informational influence (Lorenz et al., 2011; Sunstein, 2005). Also, Krause et al. (2010) showed that, in some cases, groups can even do worse than some particular individuals, for example while asked to solve problems in which the answer is highly counterintuitive. Therefore, rather than testing swarm intelligence per se, in this experiment we show its potential for solving this kind of problem, by removing any direct interaction and considering a repeatable situation where individual differences in the ability to solving the problem are measurable.

Group performance has been studied a lot, particularly in humans (Kerr and Tindale, 2004; Laughlin et al., 2002, 2006), but a problem with these previous studies is that only a single event was considered. It was therefore not possible to tell whether the individual who got the best result and beat the group performance was an “expert” at doing the task or was just being “lucky”. Experts are characterised by a consistently high performance (Shanteau et al., 2002). Our study differs from the previous ones by establishing a performance profile over several trials for each individual, enabling us to compare the performance of each individual against each other and against groups of different sizes.

A recent model predicted that in single-shot decisions, experts are almost always more accurate than the collective across a range of conditions, but that for repeated decisions – where individuals are able to consider the success of previous decision outcomes – the collective’s aggregated information is almost always superior (Katsikopoulos and King, 2010). Our study tested these predictions. We presented

(20)

18

students with a set of 10 assessment tasks to quantify differences in personal ability. To simulate a kind of problem that could be encountered by an individual in reality, we used a simple quantity estimation; i.e. estimating the number of dots on a screen. We predict that when a task is repeated several times, the group eventually outperforms even its best members.

Methods

Data collection

The experiment was carried out at the University of Applied Sciences Lübeck, Germany, in November 2008 and October 2009. All data were collected anonymously and with the permission of the participants. Students from different technical disciplines were asked to estimate the numbers of dots on 10 pictures each of which was shown on a projection screen for 10 seconds. The screen size was 160 × 120 cm and the diameter of each dot was 1.2 cm. After a picture with dots was shown the participants had 15 seconds to write down their estimate before the next picture was shown. During the time intervals a uniformly grey picture was shown.

All participants saw the same pictures in the same order. Each picture was constructed by first picking a random number n in the range 50 … 500 and then randomly placing n dots in the visible area of the screen without overlapping. To make the estimations more difficult, the dots were not uniformly placed but the probability of a position decreased with increasing distance from the centre of the screen (inset of Fig.

2.1). This prevented participants from simply counting the dots in a small subset of the area and then scaling up to obtain the total number. From the 98 answer sheets one was excluded from the evaluation because of extremely high and deliberate-looking errors.

Analysis

Each picture contained a different number of dots. In order to make errors comparable between pictures we used relative errors to measure performance. More precisely, for each individual i and picture k we defined the individual relative error d(i,k) = | g(i,k) - ck

| / ck, where g(i,k) was the guess of individual i for picture k, and ck was the correct value for picture k. In the same way we defined the collective relative error for picture k as D(k)

= | mean_i∈I g(i,k) - c_k| / c_k, where I was the set of all individuals (N = 97) and mean_i∈I

g(i,k) was the mean of the guesses of all individuals in the set I for picture k.

(21)

19

Relative error as a function of the correct value

Do the d(i,k) and D(k) depend on the magnitude of the correct values ck? It might be the case that large numbers of objects are more difficult to estimate than small ones even in terms of relative errors. This phenomenon is known as the Weber-Fechner law (Dehaene, 2003; Nieder and Dehaene, 2009). If so, the question arises whether this issue only affects the individual relative errors, or both the individual and collective relative errors. To investigate these questions we computed the mean individual relative error meani∈I d(i,k) for each picture k. Then we conducted regressions with the correct value ck as independent variable and meani∈I d(i,k) and D(k) as dependent variables.

Differences in individual ability

In order to find out whether some individuals consistently performed better than others we split the sequences of guesses in two halves, the guesses for pictures 1-5, and the guesses for pictures 6-10. For each individual i we computed the sums of relative errors for both sub-sequences, Σ_k=1…5d(i,k) and Σ_k=6…10d(i,k), and determined their correlation.

A significant correlation would be indicative of individual consistency. The d(i,k) tended to increase with increasing correct value c_k (see results section). Therefore, it seemed problematic to regard these error sums as measurements at interval scale or ratio scale level. However, the error sums certainly provide a measurement at an ordinal scale level and it made sense to use a rank correlation coefficient. Here we used Kendall’s τ_{, mainly} because it has a simple and intuitive interpretation.

Individual performance vs. collective performance

We compared individual and collective performance across (subsets of) the 10 pictures by comparing their cumulative errors. We computed the cumulative error for a subset S of the 10 pictures by summing up the errors for all pictures in S. More precisely, the cumulative error of individual i on a subset S of pictures was defined by Σ_k_∈_Sd(i,k), and the cumulative collective error on S was Σ_k_∈_SD(k). For example, if S contained the pictures 1, 4, and 5, then the cumulative error of individual i on S was d(i,1) + d(i,4) + d(i,5) and the cumulative collective error on S was D(1) + D(4) + D(5).

(22)

20

For most single pictures some individuals outperformed the group (i.e. d(i,k) <

D(k) for some individuals i and pictures k). On average, 13 individuals had a smaller error than the group for a given picture. However, regarding the cumulative error on the set of all 10 pictures no individual was better than the group. The question arose about what happens in between these extreme scenarios for intermediate numbers of estimates. To answer this question we compared the individual and the collective cumulative errors on subsets of pictures. For each subset we counted the number of individuals with smaller cumulative error than the group. In order to determine if and how this number decreases with increasing size of the subset we did this for all 2¹⁰- 1 non-empty subsets of pictures.

Additionally, to assess the influence of differences in individual abilities on the number of individuals that perform better than the group on subsets of pictures, we randomised the guesses across the individuals. More precisely, for each picture we permuted all guesses and performed the above described computations for all non-empty subsets of pictures. We repeated these steps 100,000 times to approximate the probability distribution of the number of individuals that outperform the group for each size of subsets of pictures in a null model where all individuals have the same abilities.

Statistical analyses were performed using R version 2.10.1.

Results

Relative error as a function of the correct value

The mean individual relative error increased with the magnitude of the correct value c_k (Linear regression: F_1,8= 21.15, P = 0.002, R² = 0.726; Fig. 2.1), while the collective relative error did not increase (F1,8 = 0.41, P = 0.54, R² = 0.048; Fig. 2.1). The individual relative errors increased linearly (rather than exponentially as would be expected if following a Weber-Fechner law; Fig. 2.1). The collective relative error did not increase probably because the ratio of guesses that underestimated to guesses that overestimated the correct values was roughly the same regardless of the correct value.

Differences in individual ability

Individual performance on the two halves (two sets of 5) of pictures was significantly correlated (Kendall’s tau: τ_{= 0.43, z}_τ = 6.30, P < 0.001) indicating that there were strong inter-individual differences in estimation performance. From the τ value we could

(23)

21

conclude that an individual i₁ that performed better than an individual i₂ on the first half of 5 pictures would also perform better than i₂ on the second half of 5 pictures with a probability of approximately 72%.

Individual performance vs. collective performance

The cumulative collective error on the set of all pictures was 1.08 whereas the cumulative errors of individuals were much larger and ranged from 1.73 (best performer among the 97 individuals) to 22.36. The evaluation of observed cumulative errors on subsets of pictures showed that the number of individuals that performed better than the group decreased exponentially with increasing size of the subset showing that even the strongest individual performer could not beat the group after 8 attempts (Fig. 2.2). The mean numbers extracted from the null model also decreased exponentially but were significantly smaller than the observed ones as can be seen from the 95% confidence interval (Fig. 2.2).

Figure 2.1: Mean individual error (black squares) and collective error (grey circles) on single pictures as a function of the correct value. The black bars indicate the interval that covers approximately 66% of individual errors that are smaller than the mean. We used this range rather than the standard error because of the highly asymmetric nature of the distributions. Inset: Example of the pictures shown to the participants.

0.0 0.1 0.2 0.3 0.4 0.5 0.6

0 100 200 300 400 500

Absolute value of relative error

Correct value

(24)

22

Figure 2.2: Percentage of individuals that outperform the group as a function of the size of the subset of pictures (black bars). After 8 guesses no individual in the data set was capable of beating the collective guess. For comparison the performance of the group is shown after removing individual performance differences from the data set (grey bars with 95% confidence intervals).

0%

2%

4%

6%

8%

10%

12%

14%

1 2 3 4 5 6 7 8 9

Percentage of individuals

Size of subset of pictures

(25)

23

Discussion

Our results show that, although individual relative error increases as the correct value increases, the collective relative error remains constant. One novel aspect of this study lies in the use of repeated trials by the individuals, allowing us to establish individual profiles, which showed a significant inter-individual variation. But despite this variation, the group performance beats even the best individual after 8 guesses. These results are a strong indication that the swarm intelligence potential of the group becomes particularly important when the same, or at least similar, cognitive problems are encountered more frequently. In contrast 13.4 % of individuals could beat the group for a single guess.

This result suggests that experiments in which groups were not able to beat the best individual members (Kerr and Tindale, 2004; Mathieu et al., 2008; Sunstein, 2005) used sample sizes that were either too small to provide a good collective guess or the lack of repeated trials may have given a false impression of individual quality. There are, however studies which found that even relatively small groups (of 3-5 members) can outperform their best members for some kinds of problems (Clément et al., 2013;

Laughlin et al., 2002, 2006). These are tasks in which groups achieve qualitatively better solutions to complex problems where a connection between letters and numbers need to be made and which required the development of equations. The mechanism, however, by which groups achieved this superior performance remains largely unclear. We conclude that simple estimation tasks (like the one investigated in this paper) require repeat performances to reliably assess the performance level of groups and single individuals whereas highly complex cognitive tasks can already show up a performance difference between groups and the best individuals in single trials.

What do these results tell us about the evolution of grouping via collective cognition? The results of our study indicate that the CC-benefit of group membership accumulates over time with repeat performances. Laughlin et al. (2002) suggests that an instant benefit of making even just a single decision in a relatively small group can be obtained in the case of complex problems. It would be interesting to see whether Laughlin’s approach can be modified for use with species that are known for their cognitive abilities such as non-human primates or cetaceans.

Our results indicated in our scenario how often a cognitive problem needed to occur for the group to achieve higher performance levels than the best individual. But another important consideration is of course the fitness benefit that would be associated

(26)

24

with this higher performance level. This benefit would depend on how important the cognitive problem in question is and how frequently the individuals encounter this type of problem. If the problem occurs very frequently, then even a small fitness benefit may add up over the life-time of the animals (Davies et al., 2012). However, if the problem occurs only rarely, then we should expect it to be an important one if it is to lead to the evolution of grouping via CC. Finally we also need to consider the potential costs of grouping, such as competition (Krause and Ruxton, 2002), relative to any benefits before we can predict whether factors such as CC could lead to the evolution of grouping.

Research studies on cognitive abilities in animals and humans show strong evidence for inter-individual variation (Bell et al., 2009; Dall et al., 2004; Deary et al., 2010; Healy et al., 2009; Sih et al., 2004a, 2004b). However, it remains a considerable challenge to predict the collective performance of groups that are composed of individuals of different cognitive abilities. Different outcomes are possible depending on the group composition and the information processing rules (Krause et al., 2010). This is a promising avenue for future research that takes personality differences in animals into account when looking at collective decision-making and swarm intelligence in particular.

If all individuals greatly benefit from overcoming their cognitive limitations through CC in groups, mechanisms that favour group formation and maintenance are likely to have evolved. However, what about a population in which a few (very good) individuals might not benefit greatly from being in a group (regarding decision-making) whereas the majority would? If some individuals are superior decision-makers, they might do better staying mostly alone and thereby reduce competition costs. For example, regarding more traditional grouping benefits, if living in groups has evolved because of the many-eyes effect (Krause and Ruxton, 2002), and one individual is “super aware”, then it might be expected to spend less time in groups because the costs (competition for example) are likely to outweigh the benefits (in terms of spotting predators). In the context of cognitive problem solving, we could potentially have the emergence of two strategies such as low cognitive performers that are highly gregarious and high cognitive performers that are largely solitary, provided that they have equal fitness. Potentially there could even be a strategy continuum whereby individuals of different cognitive abilities use CC to different degrees. It would then be interesting to test whether individuals with lower cognitive abilities demonstrate higher grouping tendency than individuals with higher abilities and if this relationship is affected by the type of problem encountered. Another important consideration in this context is that the development and

(27)

25

maintenance of neural tissues is very expensive (Laughlin et al., 1998). Therefore swarm intelligence might often be the energetically cheapest option to solve cognitive problems.

In this experiment, the individuals did not exchange any information while making the decision. In humans, direct interactions are not always required for swarm intelligence because, once we know how the information should be processed, we can use a computer algorithm to do the job (Krause et al., 2010; Wolfers and Zitzewitz, 2004). In fact there is some indication in the psychological literature that real interactions between individuals can be detrimental to the swarm intelligence potential of groups because of communication barriers between them (Mathieu et al., 2008; Sunstein, 2005). Therefore, at least for this range of problems, aggregating the estimates from each individual instead of allowing the members to communicate with each other is the safest solution to guarantee the quality of the decision.

One aspect we have neglected so far is the question of how the performance of the group depends on group size. Unfortunately this was beyond the scope of our study because we only had 97 individuals. To assess the performance potential of different group sizes one would need a dataset of many more individuals and their responses, to randomly draw groups of different sizes and investigate their performance relative to that of the best group member. Statistical textbooks (for example, Dorofeev and Grant, 2006) generally give the advice that a correction factor has to be taken into account when sample sizes are greater than 5% of the population. This suggests that the pool from which samples can safely be drawn for the above purpose needs to be at least 20 times larger than the largest sample size. Therefore, if we want to examine the performance of groups up to size 50, we would need a minimum of 1000 individuals in our study.

Experimental work on sparrows, Passer domesticus, showed that larger groups were more successful in solving cognitive problems (Liker and Bókony, 2009). However, they compared groups of 6 birds with those containing 2 birds and did not include single animals. The latter would be valuable in terms of understanding the evolution of grouping via cognitive benefits in groups which remains an interesting field for future research.

(28)

26

(29)

27

! Collecti e cognition in gu ies a cross- o ulation com arison study in t e wild

Clément RJG, Mann RP, Ward AJW, Kurvers RHJM, Ramnarine IW & Krause J

Collective cognition has received much attention in recent years but most of the empirical work has focused on the increase of decision-performance with group size in single populations. Here we investigated collective cognition in multiple populations that are subject to different ecological conditions. Guppies (Poecilia reticulata) were given a simultaneous choice between an edible and a non-edible stimulus. We quantified the response of fish to the test stimuli in various group sizes across four populations that differed in predation risk. Our results show that sampling activity was higher in low predation populations compared to high ones but not affected by group size or sex.

Decision accuracy increased with group size and with sampling activity of the focal individual, i.e. individuals that sampled more, and individuals in larger groups had a higher approach and peck accuracy. Group size had a significant positive effect on the probability that the first approach was made towards the edible stimulus but not on the accuracy of the first peck. Our results suggest that the use of private information (direct personal sampling) and social information (observing other group members) is context- dependent in guppies and differences in predation regimes only influenced sampling activity but not decision accuracy.

Unpublished manuscript

(30)

28

Introduction

Sociality confers many advantages to animals, such as better anti-predator defence or spotting food (Krause and Ruxton, 2002). Many of these advantages result from collective cognition, which is the ability of members of groups to make decisions that are better than those made by single individuals (Krause et al., 2010).

Animals often gain advantage from paying attention to cues that are inadvertently displayed by their conspecifics, using this social information to complement their own information in order to make better decisions (Danchin et al., 2004). For example, rats infer from their congeners’ breath whether some food is safe or not to consume (Galef, 1991), whereas starlings observe foraging success of their flockmates to assess the quality of a food patch (Templeton and Giraldeau, 1995). Social information is also used by social animals to make collective decisions based on quorum thresholds. For instance shoals of stickleback (Gasterosteus aculeatus) use quorum thresholds to evaluate predation risk (Ward et al., 2008) or to locate foraging patches (Ward et al., 2012).

Similarly, ant colonies that are forced to emigrate from their nest are able to choose the best available new nest by accelerating the recruitment process once a threshold number of ants have made a decision in favour of a particular location (Pratt et al., 2002; Sumpter and Pratt, 2009). It was also shown in fish shoals that the speed and the accuracy of decisions increased with group size (Ward et al., 2011).

One important selection pressure affecting the use of social information is thought to be predation risk (Devereux et al., 2006; Elgar, 1989). For example, experimentally increasing the perceived risk of predation in the lab resulted in an increased reliance on social information in Minnows (Phoxinus phoxinus) due to increased cost of gathering private information (Webster and Laland, 2008). Three-spined sticklebacks (Gasterosteus aculeatus) have a thick body armour making it less risky to collect private information, as compared to the sympatrically occurring nine-spined sticklebacks (Pungitius pungitius), which lack this armour and are more vulnerable to predation and more prone to using social information (van Bergen et al., 2004; Coolen et al., 2003, 2005). Furthermore, it has been shown that predation risk can influence social structure of animal groups (Kelley et al., 2011), possibly affecting information transfer, and therefore social learning.

However, a largely unresolved question is how individuals from populations under different predation regimes in the wild differ in their use of private versus social information.

(31)

29

Many fish species feed on objects falling into the water, and individuals have to respond quickly to a stimulus and consume it before others do (Krause, 1993). However, not every fallen object is edible and the fish will often approach objects that they cannot consume at the cost of wasting energy, missing simultaneously occurring and genuine feeding opportunities, and exposing themselves to predation (Hall et al., 2013; Ioannou and Krause, 2009).

In many natural situations, timing matters, and little differences can lead to different outcomes among conspecifics. For example, detecting and responding to a predator earlier increases the chance of survival, while predators focus on individuals that react more slowly (Kenward, 1978; Post et al., 2013). Similarly, in a foraging context, being part of a group implies competing with other members when the food is limited and a rapid response to opportunities can increase food intake. For example one might benefit from social information about a fallen object by watching how the others react to it, but a delayed response may mean a missed opportunity. Authors previously looked at single populations in the context of decision accuracy (Sumpter et al., 2008a; Ward et al., 2008, 2012). Here we used the comparative approach to investigate groups of free-ranging guppies from four different populations facing different predation regimes to study the effects of predation risk on the evolution of collective cognition.

We studied the effects of group size and predation level on decision accuracy by presenting fish simultaneously with an edible and a (similar looking) non-edible item. We quantified how frequently the edible and the non-edible stimuli were targeted as a measure of decision accuracy. Furthermore, we quantified general sampling activity of individuals as a proxy of their level of private information. We studied how group size and predation level affected this sampling activity, and how sampling activity, in turn, affected decision accuracy.

Carotenoid pigments have been shown to confer health benefits (Kolluru et al., 2006). They are sequestered for use in courtship displays by male guppies (Kodric- Brown, 1989), thus also conferring fitness benefits. They cannot be synthesised by guppies and must be obtained through their diet (Fox, 1976). They are present in orange and red fruits that are abundant in the rainforest and often end up in the streambed, where they are very attractive to guppies, both males and females (Rodd et al., 2002). However, at the same time the streambed usually contains many other reddish objects including pebbles that are non-edible. We therefore presented edible and non-edible stimuli made

(32)

30

out of red materials that resemble each other to single fish and shoals of various sizes and evaluated the decision accuracy as a function of group size.

Due to the anti-predator effects of group-living, sampling is usually safer for individuals in larger groups (Krause and Ruxton, 2002). We therefore predicted that i) sampling activity would be positively correlated with group size. Given that sampling should be safer in lower predation sites we also predicted that ii) sampling activity is greater in low predation sites than in high predation sites. Finally, we expected that iii) decision accuracy would be positively correlated with group size, as predicted by previous work (Ward et al., 2008, 2011).

Methods

Experimental setup

The study took place in the Turure River (lower: 10°39’27”N, 61°9’48”W; upper:

10°41’7”N, 61°10’23”W) and the Aripo River (lower: 10°39’1”N, 61°13’26”W; upper:

10°40’55”N, 61°13’51”W) in Trinidad (March 2013, 2014 & 2015). These rivers consist of interconnected pools inhabited by different-sized guppy populations (Poecilia reticulata). Both streams are known to have a sharp gradient in predation pressure: in the lower sections (below the main falls) characids and cichlids are present, which heavily predate on guppies. These predators are absent in the upper sections (above the main falls) (Magurran, 2005). This provides a unique opportunity to use a natural gradient in predation pressure to investigate its effect on collective cognition. We sampled populations from below and above the main falls in both rivers.

Individuals from our target populations are known to respond strongly to orange and red items falling on the water surface (usually these are edible fruits dropping from trees into the stream). Therefore, an edible stimulus and a non-edible stimulus were made out of red material. The edible stimulus was a piece of red bell pepper and the non-edible stimulus was cut out of red plastic. Both stimuli had the same shape (9*5*2 mm) and each was fixed to a weighted monofilament line (40 cm long, ø 0.2 mm) attached to the end of a wooden rod, 20 cm apart from each other. We first conducted pilot trials to verify that both stimuli were effective with fish in the wild when presented in isolation (mean ± SE number of approaches towards the edible stimulus: 4.1 ± 0.5 (n = 30), towards the non-edible stimulus: 2.6 ± 0.4 (n = 37); mean ± SE number of pecks at the edible stimulus: 3.7 ± 1.0, at the non-edible stimulus: 1.6 ± 0.5).

(33)

31

Before the start of each trial we counted the number of guppies present in the selected group and randomly selected one individual as focal individual (because often, the fish would travel beyond the field of the camera, making problematic the recording of entire groups) of which we determined sex and size class (small/middle/large). Both stimuli were then slowly and simultaneously lowered into the water approximately equidistant from the focal individual (and whenever possible also equidistant from the whole group). Trials lasted 2 minutes following the introduction of the stimuli, giving the fish sufficient time to inspect the stimuli (although our main results did not change when considering shorter time periods (down to 30 s)). We scored the number of approaches and pecks made by the focal individual towards the edible and the non-edible stimuli. An approach was defined as the focal individual moving towards the stimulus within one body length from it and a peck was defined as the focal individual biting or nibbling the stimulus. A peck always followed either an approach or a previous peck whereas an approach was not necessarily followed by a peck. In total 607 trials were conducted, testing fish in group sizes ranging from 1 to 72 (mean group size = 6.2). Each trial was carried out in a different location along both streams and it is thus highly unlikely that the same groups and the same individuals were tested twice.

Analysis

We first investigated the effect of group size and predation level on sampling activity. For each randomly selected individual, we quantified every approach and peck toward each stimulus. To account for the high number of zeros in our data, we looked at the sampling activity using a hurdle model (count model with truncated negative binomial distribution and log link, zero hurdle model with binomial distribution and logit link) with group size, predation level, interaction between group size and predation level, river, size and sex as explanatory variables. We used likelihood ratio tests to compare the different models. We ran different models for approaches and pecks as they reflect qualitatively different decision processes despite both indicating a level of interest towards the stimuli.

We then studied decision accuracy, using the ratio (edible / (edible + non-edible)) as a measure of decision accuracy separating again approaches and pecks. We used this ratio as response variable in a generalized linear model with sampling activity (sum of approaches and pecks), group size, interaction between activity and group size, predation

(34)

32

level, river, size and sex as explanatory variables. We used binomial errors and a logit- link function since the response variable was bound between zero and one (with 0 indicating that all approaches/pecks were towards the non-edible stimulus, and 1 indicating that all approaches/pecks were towards the edible stimulus).

Finally, to minimize the effect of sampling activity and focus on the effect of group size, we also looked at the first decision made (i.e. first approach and first peck).

Results

Sampling activity

The number of approaches was not affected by group size (Fig. 3.1; count model: ² = 0.02, df = 1, p = 0.89). However, there was a significant effect of predation level (Fig.

3.1; count model: ² = 7.15, df = 1, p = 0.008; zero hurdle model: ² = 2.66, df = 1, p = 0.10), with fish in low predation areas approaching both types of stimuli more often than fish in high predation area. There was no significant interaction effect between group size and predation level (count model: ² = 0.22, df = 2, p = 0.90). There was also a significant difference between the two rivers (Fig. 3.1; count model: ² = 12.52, df = 1, p

< 0.001; zero hurdle model: ² = 11.41, df = 1, p < 0.001), with fish in the Turure River being more likely to approach, and approaching more often than fish in the Aripo River.

There was no effect of sex (count model: ² = 1.23, df = 1, p = 0.27) but there was an effect of the size of focal fish (Fig. 3.1; count model: ² = 4.31, df = 1, p = 0.032; zero hurdle model: ² = 7.27, df = 1, p = 0.007), with larger fish being more likely to approach both types of stimuli, and approaching them more often than smaller fish.

Group size did also not affect the number of pecks (Fig. 3.1; count model: ² = 1.08, df = 1, p = 0.30). There was, again, a significant effect of predation level (Fig. 3.1;

count model: ² = 14.54, df = 1, p < 0.001; zero hurdle model: ² = 3.42, df = 1, p = 0.064), with fish in low predation areas pecking at both types of stimuli more often than fish in high predation areas. There was also no significant interaction effect between group size and predation level (count model: ² = 1.60, df = 2, p = 0.45). There was no significant difference between the two rivers (Fig. 3.1; count model: ² = 2.70, df = , p = 0.10). There was no effect of sex (count model: ² = 0.96, df = 1, p = 0.33) but there was an effect of the size of focal fish (Fig. 3.1; count model: ² = 5.01, df = 1, p = 0.025; zero

(35)

33

hurdle model: ² = 29.13, df = 1, p < 0.001), with larger fish being more likely to peck, and pecking more often than smaller fish.

Figure 3.1: Sampling activity is greater in low predation sites (upper parts of the rivers) than in high predation sites (lower parts of rivers). It is also greater for larger individuals than smaller ones. Additionally, it is greater in the Turure River than in the Aripo River.

Shown are medians and interquartile ranges. Data points outside 1.5 times the interquartile ranges are displayed separately.

Collective cognition and decision-making in humans and fish